Virtual Library

Start Your Search

Yunlang She



Author of

  • +

    MA03 - Lung Cancer Screening - Next Step (ID 896)

    • Event: WCLC 2018
    • Type: Mini Oral Abstract Session
    • Track: Screening and Early Detection
    • Presentations: 1
    • Moderators:
    • Coordinates: 9/24/2018, 10:30 - 12:00, Room 206 AC
    • +

      MA03.07 - Development and Validation of Deep Learning ModelĀ for Recognition of Histologic Subtype of Lung Adenocarcinoma from CT Images (ID 14412)

      11:10 - 11:15  |  Presenting Author(s): Yunlang She

      • Abstract
      • Presentation
      • Slides

      Background

      The clinical decision to either follow-up or resection from radiologic features for lung adenocarcinoma (atypical adenomatous hyperplasia [AAH], adenocarcinoma in situ [AIS], minimally invasive adenocarcinoma [MIA] and invasive adenocarcinoma [IA]) appearing as Sub-solid nodules (SSNs) is still challenge, and currently more relies on measures of diameter, solid component ratio. With the successful application of deep learning neuro-network (DLNN) for the classification of skin or common treatable blinding retinal diseases, we hypothesized that DLNN might help the histologic subtype classification of SSNs from CT images. The purpose of this study is to develop and validate a deep neuro-network model to classify AAH, AIS, MIA and IA or define a feasible classification for follow-up or treatment decision.

      a9ded1e5ce5d75814730bb4caaf49419 Method

      A total of 869 patients with 1344 pathologic confirmed nodules (AAH: 75, AIS: 340, MIA:321, IA: 608) were enrolled into this study. Two 3D mixed-scale dense-connected convolutional neuro network models (3D MS-DenseNet) were developed for 2 classification tasks: 4-class (AAH, AIS, MIA, IA), 3-class (AAH, AIS/MIA, IA). Eighty percent of whole datasets were randomly selected for training set, while other 20% were used for testing set. The nodules were firstly selected using a bounding box in 3D Slicer, and then cropped into 128 x 128 x 128 matrix size as the input to MS-DenseNet, and the output layer from the network was a 4-node or 3-node softmax classifier. Confusion matrix were used for the performance evaluation of both models and the classification accuracy for each class were reported.

      4c3880bb027f159e801041b1021e88e8 Result

      The classification accuracy of AAH, AIS, MIA, IA in testing set were 0.75, 0.45, 0.52, 0.85 respectively by 4-class, suggesting that the differentiation between AIS and MIA from CT images by neuro-network is challenge. While in the 3-class classification task with purpose of decision supporting for treatment, the classification accuracy of AAH, AIS/MIA, IA were 0.70, 0.73, 0.88 in the same testing set.

      8eea62084ca7e541d918e823422bd82e Conclusion

      The DLNN showed potential capability in differentiating AAH, IA from other adenocarcinoma subtypes, while failed to differentiate AIS and MIA. When combing AIS and MIA for reclassify adenocarcinoma subtypes from the perspective of treatment, the DLNN achieved reasonable performance, suggesting that DLNN might be useful in supporting clinical treatment decision whether to follow-up or take different resection for SSNs.

      6f8b794f3246b0c1e1780bb4d4d5dc53

      Only Members that have purchased this event or have registered via an access code will be able to view this content. To view this presentation, please login, select "Add to Cart" and proceed to checkout. If you would like to become a member of IASLC, please click here.

      Only Active Members that have purchased this event or have registered via an access code will be able to view this content. To view this presentation, please login or select "Add to Cart" and proceed to checkout.